Deepfakes Deconstructed đź”®

A New Era of Threat and Opportunity in Digital Media.

By: Evan Luksay

TL;DR

Synthetic visual or auditory media developed utilizing advanced deep learning techniques — dubbed Deepfakes — have become nearly indistinguishable from reality, presenting enormous implications, particularly within social and political contexts.

The Breakdown

  • The term Deepfake — arising from its production of fake media through deep learning technology — emerged in 2017 through viral videos showing seemingly “real” public figures spouting blatantly fictitious rhetoric
  • Thanks to technological innovation within AI and Computer Vision, hyper-realistic CGI is no longer reserved to billion-dollar animation companies like Pixar and DreamWorks
  • From photoshopped images, to fictitious lip-synching, to synthetic speech to complete physical replication, Deepfakes provide an unprecedented number of channels through which everyday people can produce counterfeit media
  • Due to the chilling precision of these synthetic fabrications, experts are wary of individuals using the technology with ill intent, prompting the FBI to issue a public warning

The Tech

  • Deepfakes rely on Deep Learning — a branch of Machine Learning — characterized by the use of neural networks loosely mimicking the functionality of human brains algorithmically
  • A type of neural network called an autoencoder is used to produce Deepfakes. An autoencoder is comprised of 1) an encoder that first reduces an image to a lower-dimensional latent space (a representation of compressed data) and 2) a decoder that reconstructs the image from the aforementioned latent space
  • Through the process of deconstructing and reconstructing the image, the autoencoder is effectively trained in an unsupervised manner to learn the composition of an individual’s face
  • In many cases, a second neural network, called a generative adversarial network (GAN), is used to perfect the Deepfake
  • GANs work by having two contesting neural networks: 1) a generator that generates synthetic media from a sample and 2) a discriminator that has been trained to classify media as real or fake
  • The discriminator gives real-time feedback via a realism score to the generator, thereby causing the generator to actively change the neural network’s weighting to produce a more realistic Deepfake

The Significance

  • Legitimate applications for Deepfake technology do exist, and include movie production, digital art, criminal forensics, and even engaging historical reenactments that aid in teachers’ lessons within the educational system
  • Nonetheless, there are still grave concerns that Deepfake technology will continue to be harnessed with malicious intent by a subset of users. Most notably is the alarming quantity of Deepfake pornography, particularly aimed against females. In response, some governments have swiftly introduced legislation making this practice illegal
  • In the age of social media, fake videos could easily be used to blackmail individuals, or at a geopolitical level, for propagandistic purposes — whereby the implications are that much more significant and potentially dangerous
  • Fortunately, AI researchers and governments are developing software to detect these fabrications. One such program has already achieved 94% accuracy in detecting Deepfakes by scanning the fabricated individual’s eyes

An exploration of the most cutting-edge applications of AI in short, insightful articles published weekly.

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store